Deep spectral component filtering as a foundation model for spectral analysis demonstrated in metabolic profiling

Abstract

Analysing metabolites in bioliquids through various spectroscopic methods provides valuable insights into the metabolic phenotypes. Deciphering spectral data has greatly benefited from deep-learning methods; however, data-driven solutions often struggle with data dependence on different devices, samples and spectral modalities. Most current task-specific methods have limited generalizability to different spectral analysis problems, including preprocessing, quantification and interpretation. Here, we developed a pretrained foundation model, termed deep-spectral component filtering (DSCF) through a self-supervised approach termed spectral component resolvable learning. By acquiring general spectral knowledge, DSCF achieved state-of-the-art performance for five distinct spectral analysis tasks on 11 datasets. Notably, the general pretraining led to zero-shot spectral denoising and trace-level quantification in complex mixtures. DSCF achieved molecule-level interpretation of surface-enhanced Raman spectra and mapped serum metabolic profiles from nearly 600 individuals for various diseases, including stroke, Alzheimer’s disease and prostate cancer. Overall, the proposed foundation model illustrates promising generalizability for spectral analysis and offers a clear and feasible pathway for general spectral analysis.